Comparing Models of Human Path Integration

Author: Fabian Kessler

Supervisor: Prof. Constantin Rothkopf, PhD. & Dr. Julia Frankenstein

Submission: Nov. 2019

Abstract:

The ability to navigate and orient ourselves in the world is crucial for our everyday lives. Humans combine many different cues in order to reduce their navigational variability. In the absence of landmarks, humans solely rely on self-motion information derived from multiple sensory sources and integrate them over time. This strategy is known as path integration.

Solely relying on path integration leads to large behavioral biases over temporal and spatial extents. In consequence, biases in path integration have been observed across many species of mammals and humans in a variety of experimental tasks. Previously, these biases have been attributed to leaky integration over the extent of time or space traveled. From a Bayesian viewpoint, leaky integration has been considered a severely sub-optimal strategy, and path integration errors are better understood as the optimal solution of the problem given noisy sensory input in combination with task-specific goals.

Two recent Bayesian path integration models, attribute biases in (visual) path integration to two different sensory priors based on the environmental statistics. Either about target distances (Petzschner & Glasauer, 2011), or slow-velocities (Lakshminarasimhan et al., 2018). Based on a comparison between the two studies, a set of experiments was derived that aimed to clarify how the two sensory priors interact.

Subjects traveled toward previously perceived target positions under varying levels of speed and self-motion uncertainty. Results show that biases in path integration behavior were strongly influenced by differences in the initial perception of the target, the reliability of self-motion cues, and the duration of travel. Subjects response variability was shown to scale linearly with distance following Weber’s Law. A comparison of three path integration models yielded that neither of the concurrent models can fully account for the variability in subjects’ distance responses. The observed variability can be used to formulate a more complete model, but further experiments are required to lay the groundwork for such a model.